Small Area Social Indicators for the Indigenous Population: Synthetic data methodology for creating small area estimates of Indigenous disadvantage

Transcription

1 Small Area Social Indicators for the Indigenous Population: Synthetic data methodology for creating small area estimates of Indigenous disadvantage NATSEM Working Paper 13/24 Yogi Vidyattama Robert Tanton Nicholas Biddle September 2012 i

2 About NATSEM The National Centre for Social and Economic Modelling was established on 1 January 1993, and supports its activities through research grants, commissioned research and longer term contracts for model maintenance and development. NATSEM aims to be a key contributor to social and economic policy debate and analysis by developing models of the highest quality, undertaking independent and impartial research, and supplying valued consultancy services. Policy changes often have to be made without sufficient information about either the current environment or the consequences of change. NATSEM specialises in analysing data and producing models so that decision makers have the best possible quantitative information on which to base their decisions. NATSEM has an international reputation as a centre of excellence for analysing microdata and constructing microsimulation models. Such data and models commence with the records of real (but unidentifiable) Australians. Analysis typically begins by looking at either the characteristics or the impact of a policy change on an individual household, building up to the bigger picture by looking at many individual cases through the use of large datasets. It must be emphasised that NATSEM does not have views on policy. All opinions are the authors own and are not necessarily shared by NATSEM. NATSEM, University of Canberra 2013 All rights reserved. Apart from fair dealing for the purposes of research or private study, or criticism or review, as permitted under the Copyright Act 1968, no part of this publication may be reproduced, stored or transmitted in any form or by any means without the prior permission in writing of the publisher. National Centre for Social and Economic Modelling University of Canberra ACT 2601 Australia Phone Fax natsem@natsem.canberra.edu.au Website ii

3 Contents About NATSEM Author note General caveat Abstract ii iv iv v Introduction 6 Data and Methodology Data Sources Methodology The variables The variables of interest Benchmarking variables between the Census and NATSISS 13 Building the Synthetic Unit Record data Creation of the unit record data based on Census tables Imputing the NATSISS variables Independent variables for the regression Regression statistics Overall imputation result 31 Applying the spatial microsimulation model Benchmark tables Results and validation from the spatial microsimulation 35 Concluding Remarks 41 References 41 iii

4 Author note Robert Tanton is Research Director of the Regional and Urban Modelling team at NATSEM at the University of Canberra. Yogi Vidyattama is a Senior Research Fellow at NATSEM and part of the Regional and Urban Modelling. Nicholas Biddle is an applied behavioural scientist and a Fellow at the Centre for Aboriginal Economic Policy Research (CAEPR) at the Australian National University (ANU). General caveat NATSEM research findings are generally based on estimated characteristics of the population. Such estimates are usually derived from the application of microsimulation modelling techniques to microdata based on sample surveys. These estimates may be different from the actual characteristics of the population because of sampling and nonsampling errors in the microdata and because of the assumptions underlying the modelling techniques. The microdata do not contain any information that enables identification of the individuals or families to which they refer. The citation for this paper is: Vidyattama,Y., Tanton,R., and Biddle, N. (2013), Small Area Social Indicators for the Indigenous Population: Synthetic data methodology for creating small area estimates of Indigenous disadvantage, NATSEM Working Paper 2013/24, NATSEM: Canberra iv

5 Abstract The lack of data on how the social condition of Indigenous people varies throughout Australia has created difficulties in allocating government and community programs across Indigenous communities. In the past, spatial microsimulation has been used to derive small area estimates to overcome such difficulties. However, for previous applications, a record unit file from a survey dataset has always been available on which to conduct the spatial microsimulation. For the case of indigenous disadvantage, this record unit file was not available due to the scarcity of the Indigenous population in Australia, and concerns from the ABS about confidentialising the file. This study offers a solution to this problem by proposing the building of a synthetic unit record file with observations that sum to the population totals from the actual survey file. A spatial microsimulation approach is then applied to this synthetic unit record file and the results are validated. v

6 1 Introduction A key input into the development of public policy in Australia is the geographic distribution of socioeconomic outcomes. Evidence-based assessment of need has to take into account not only where people live, but also their characteristics. This has led to the widespread use of the Socio-economic Indexes for Areas (SEIFA), developed by the ABS and updated each five-yearly census. However, these indexes do not capture the differences in Indigenous disadvantage across the country. Evidence has shown that an index of Indigenous disadvantage gives different results to a general index of disadvantage like SEIFA, primarily due to the fact that non-indigenous advantage masks a great deal of Indigenous disadvantage (see Kennedy and Firman, 2004). Furthermore, there are different components of wellbeing that need to be considered for Indigenous people that are not captured by the Census. The need for different social indicators at a local level has caused small area estimation techniques to flourish (Ghosh and Rao, 1994; Pfeffermann, 2002). These techniques include methodologies that are based on the regression of predictors (Elbers et al 2003, Pratesi and Salvati 2008, Fabrizi et al 2012) and the reweighting of survey data known as spatial microsimulation modelling (Caldwell et al., 1998; Ballas et al., 2005 Nakaya et al., 2007; Vidyattama et al., 2011). In Australia, the spatial microsimulation method has increasingly been used in the past to derive small area estimates of a range of economic and social indicators, as well as estimating the impact of government policy and the need for government services at a small area level (Harding and Tanton, 2011). Examples of this work include simulating the small area impact of changes in income taxes and cash transfers (Chin et al., 2005; Harding et al., 2009; Tanton et al., 2009); the need for different types of aged care (Lymer et al., 2008); the retirement saving by gender of those who have just retired (Vidyattama et al., 2011) as well as measuring the distribution of trust (Hermes and Poulsen, 2012). All of these small area estimation techniques bring together survey data that contains a specific variable but does not have enough observations to represent a small area with census or administrative data with enough observations in a small area to derive a reliable estimate. This is especially true for spatial microsimulation methods that require a unit record survey file. Unfortunately, in some cases, this unit record survey file is not available, maybe due to confidentiality reasons; or because the owner of the dataset cannot release it. This is the case with the National Aboriginal and Torres Strait Islander Social Survey (NATSISS) from the Australian Bureau of Statistics (ABS). This file contains information that is considered sensitive by the ABS and is therefore not released to the public. This work reports on the development of a method that allows spatial microsimulation to be conducted when survey unit record data is not available. The model used for this paper is one variant of a spatial microsimulation model that uses a generalised regression reweighting method to reweight survey data to known small area totals from the Census. 6

7 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. Following the introduction, the structure of this paper is as follows. Section two presents the data and methodology that will be used to construct the data base. Section three goes through the process of building the synthetic unit record data; Section four describes the application of the spatial microsimulation process; and Section five concludes the paper. 2 Data and Methodology 2.1 Data Sources As mentioned in the introduction, the estimation process requires two main databases. The first is the census of population and housing. The Australian Bureau of Statistics (ABS) conducts a nationwide Census to obtain a count of the number of people in Australia, their individual characteristics and their dwelling characteristics every five years. The latest two censuses were on 8 August 2006 and 9 August 2011, respectively. Both capture more than 20 million people. This study used those two census datasets. Indigenous people in the census are identified by the question that asks whether a person is of Aboriginal and/or Torres Strait Islander decent. Among the 20 million people in Australia, around 455,000 in 2006 and 548,000 in 2011 identified themselves as Aboriginal or Torres Strait Islander. Between the 2006 and 2011 census, the increase in the number of Indigenous people was 4.10 per cent annually, higher than the total population. The high growth rate in the number of Indigenous people does not come only from population growth, but may also be due to an improvement in Indigenous enumeration and selfidentification following the ABS Indigenous Enumeration Strategy (IES). An evaluation conducted of the IES after the 2006 census has led to continuous and extended Indigenous community engagement and involvement in the 2011 census (ABS 2011a, Morphy et al., 2007). The census data in Australia contain characteristics such as age and sex, cultural and language diversity, disability and carer status, children and childcare, employment and income, education and qualifications as well as relationship. Some of the questions in the census also ask about the condition of the family and the household such as the number of children, family composition, dwelling characteristic and tenure and household income. Most of the questions are multiple choice with few open-ended questions. Some of the numeric data such as income are collected based on income ranges. The main advantage of using census data for this study is the geographical information it contains. Two different Census s were used for this work. The synthetic Indigenous dataset used the 2006 Census data, while the spatial microsimulation process used the 2011 Census data. The reason for this was that the 2011 Census data was not available when the synthetic Indigenous dataset was created. The ABS currently uses the Australian Statistical Geography Standard (ASGS) as the main structure of census data dissemination. This has replaced the use of the ASGC since 2011.In the ASGS, Statistical Areas Level 1 (SA1s) are designed to have an average population of about 400 while Statistical Areas Level 2 (SA2s) have an average population of about 10,000, with a minimum population of 3,000 and a maximum of 25,000. 7

8 Besides the ASGS classification, there are additional area classifications in the census based on other known structures. The Indigenous Structure is one of these. This area is based on the existence of known Indigenous communities that should be represented by the area. The second database used for this work is the National Aboriginal and Torres Strait Islander Social Survey (NATSISS) The main aim of this survey is to capture health, education, culture and labour force participation of Indigenous people. In addition, the 2008 survey also captured population characteristics such as age and sex, social capital such as social contact and networks, life experiences such as bullying and discrimination, housing and mobility, transport, information technology (i.e. computer, internet, telephone) as well as safety and experiences of crime in the home. The ABS conducted the 2008 survey using face to face interviews between August 2008 and April There were 13,307 observations in the 2008 survey to represent around 518,000 Indigenous people across Australia. Although most of the Indigenous population lived in either New South Wales (around 154,000) or Queensland (around 146,000), there was a disproportionately high number of observations from Victoria and the Northern Territory with 2,245 and 2,267 observations, respectively. The high number of observations in the Northern Territory is due to the fact that Indigenous people in the Northern Territory tend to be in smaller and more heterogeneous communities requiring a greater sample from each community to derive accurate results for each area (ABS 2008). The main geographic record in the NATSISS is the State and Territory with the Australian Capital Territory (ACT) and Tasmania amalgamated. However, given around 70 per cent of Indigenous people lived outside the capital cities, we analysed the areas outside the capital cities based on the ABS remoteness structure which was available on the NATSISS. The remoteness structure is developed based on the road distance of an area to several closest service centres as indicated by population. With the remoteness structure, the ABS has created 17 areas with Queensland divided into four defined areas (Major Cities, Inner Regional, Outer Regional and Remote/Very Remote areas), New South Wales into three defined areas (Major Cities, Inner Regional, and Outer Regional), Victoria into two (Major Cities and Inner/Outer Regional) and Western Australia into two areas (Non-Remote and Remote/Very Remote). There are specific areas for the non-remote areas of South Australia and Tasmania as well as the remote area of the Northern Territory while the other three states/territories and the Australian Capital Territory are grouped as part of Balance of Australia. Given the sensitivity of the data, the ABS has introduced a special arrangement for the dissemination of this survey. While the summary results from the 2008 NATSISS are available in a national level publication (see ABS 2010), the unit record data of NATSISS is not available to researchers. However, it is possible to have authorised access to the Confidentialised Unit Record File (CURF) through the ABS online query system by submitting code to the ABS Remote Access Data Laboratory (RADL) (ABS, 2006). The ABS can reject a request if it results in cells with a low number of observations or when the statistical or econometrical procedure uses a very specific sub-sample of the survey. 2.2 Methodology Spatial microsimulation has emerged as a well-established technique in the estimation of small area statistics. Spatial microsimulation uses the individual or household units from 8

9 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. survey data to populate each small area, subject to constraints from Census tables. These Census tables have a number of different classes in each table (for instance, there will be an estimate for the number of males aged who are unemployed in a particular geographic area). These benchmark classes in each benchmark table give information about the distribution at the smallest spatial area available in the data (Williamson et al., 1998; Voas and Williamson, 2000; Williamson, 2001). Reweighting techniques have become one of the most common approaches to the creation of synthetic spatial microdata and, within this broad method, there are a number of different methodologies available (Tanton and Edwards, 2013; Anderson, 2007; Ballas, et al., 2005; Hynes, et al., 2009; Tanton et al., 2011; van Leeuwen et al., 2009; Voas and Williamson, 2000, Zaidi et al., 2009, Rahman et al., 2010). The SpatialMSM model used for this paper employs a generalised regression reweighting program from the Australian Bureau of Statistics (ABS) called GREGWT. The GREGWT algorithm uses a generalised regression technique to estimate weights for a household or individual from the survey, and then iterates until the weighted aggregate of the survey data produces characteristics that closely resemble the constraints for each small area (Bell, 2000; Tanton et al., 2011). The procedure can be classified as a deterministic method using formulae, similar to the iterative proportional fitting used by Anderson (2007) and Ballas et al. (2005), as opposed to a probabilistic method that pseudo-randomly selects households to fill an area described in other models (Voas and Williamson, 2000; Williamson et al., 1998). Despite the differences in approach, Tanton et al. (2007) confirms that the results from different reweighting methods are generally similar. In Australia, spatial microsimulation has used unit record files from a number of surveys including the Survey of Income and Housing (SIH), the Household Expenditure Survey (HES) and The Household, Income and Labour Dynamics in Australia (HILDA). Given the unit record data from NATSISS survey was not available for this study, another set of synthetic data was created using two stages of an imputation method. Figure 1 shows the flow chart of the estimation process including the two stages of imputation used to create the synthetic database. In the first stage, we imputed the data using a probabilistic table from Census data as in Williamson (2013). The main reason for using the probabilistic table was that the Tablebuilder program used to extract the Census data from the ABS allows us to create cross tabulations that can contain up to around 5,000,000 cells. This means we may create a set of probability tables with around four conditional factors. As shown in Figure 1, this imputation would create a synthetic database made up of census data only. However, we also need the synthetic database to contain the variables that we would like to estimate from NATSISS. The probabilistic table technique cannot be used for this as the construction of the table will require the use of a relatively small number of observations to fill the cells of a cross tabulation, even with a small number of conditioning factors. Therefore, the second stage of the imputation process uses a regression method to impute the specific conditions that are available from the NATSISS detailed State by Remoteness (ABS 2008) table onto the synthetic database. Regression on variables of interest from the NATSISS will produce the coefficients needed to impute the variables onto the synthetic dataset. Given most of the variables of interest are likely to be binomial, the variables are estimated using a logit or probit regression model and the application of the coefficient will allow us to find the probability of the condition for each observation. A 9

10 random number can then be applied to estimate the binomial value. Once all the necessary variables from NATSISS were successfully imputed, we can begin the reweighting process. Figure 1 Flowchart of the estimation procedure The database with the number of observation equal to total ATSI population divided by the pre-set weight 14 Census probability tables for different major statistical region Imputation of data based on Census probabilistic table NATSISS data on ABS RADL national level Indigenous synthetic dataset with census variables Regression of the variables of interest with common variables Imputation of data using the coefficients estimate from NATSISS Regression Coefficient estimates national level Indigenous synthetic dataset with census and NATSISS variables 2011 census Benchmark tables for Small area Reweighting in SpatialMSM Indigenous synthetic dataset with small area weights Note that there are some areas where an estimate cannot be produced by this reweighting process. This is mostly because the process does not achieve an acceptable error for the estimate. The error in the reweighting process is measured by the total absolute error (TAE) from all the benchmarks. The TAE is calculated by summing all the differences between the 10

11 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. estimated number from the model and the benchmark number from every benchmark class of each benchmark table. The total error threshold that is set for this spatial microsimulation model is whether the TAE from all the benchmarks is greater than the population in that area. The TAE has been used in a number of spatial microsimulation models as a criterion for reweighting accuracy (Anderson, 2007; Williamson, et al., 1998) and has been assessed and supported by other studies such as Smith et al. (2009) and Voas and Williamson (2000). 2.3 The variables This project aims to develop specific measures of wellbeing for the Indigenous population. The indicators chosen after some consultation with experts on Indigenous wellbeing are: participation in cultural activities; social capital; discrimination; health status; psychological stress; social and emotional wellbeing; financial stress; feelings of safety or stress; identification with clan, tribal or language group. As highlighted in Cassells et al., (2010), the spatial microsimulation process starts by identifying relevant variables that exist in both databases. The identification of these variables in this study is not only important for the reweighting process but also for the imputation of data through both the probabilistic table and the regression process. The accuracy of spatial microsimulation can only be ensured when benchmarking variables are correlated with the variables being estimated from the model (Anderson, 2007) The variables of interest (a) Participation in selected cultural activities in last 12 months This variable is available on the NATSISS as WCULACT and available for persons aged 15 years and above. This binary variable relates to the specific activities that have a cultural association including fishing, hunting, gathering wild plants / berries, making Aboriginal and Torres Strait Island arts or crafts, performing any Aboriginal and Torres Strait Island music, dance, theatre or writing or telling any Aboriginal and Torres Strait Island stories. (b) Social capital This variable is an index from -1 to 1 that was derived using Principal Component Analysis from several social capital variables including social cultural aspects such as involvement and attendance in cultural events, ceremonies or organisations (WCULEVNT andculpq12) and in sporting, social or community activities (WPAR3M). Social capital also considers the relationship to family and friends such as the frequency of contact with family or friends outside the household (FACCONC, NFACCOC and FANYCOC), whether the person can confide in family and friends(wfrndcon) as well as the proportion of friends of the same 11

12 age, Indigenous origin or similar education (PFRNDAGE, PFRNDIND, PFRNDEDU). Social capital also includes the ability to get or to give support outside the household (FCSPQ1, WSPTREL, WSPTANY), social efficacy (comfortable contributing in the family and community) and trust (in general and to the medical profession or police). (c) Felt discriminated against in last 12 months This binary variable WDISC12 is the identification of discrimination in general. This includes discrimination at the workplace, at the neighbourhood and at educational institutions as well as while doing any sporting, recreational or leisure activities. The discrimination can also relate to treatment by the police, security guards, lawyers, in a court of law, by medical apparatus, or by any staff of Government agencies when seeking any public services. (d) Poor health status This variable is based on self-assessed health (SAHQ1) where the person was asked to identify his or her health using the following categories: excellent, very good, good, fair and poor. There were more than 75 per cent of respondents rated their health as good or better, and around 10 per cent reported they had poor health. For this variable, we have used a binary response of feeling in poor health or not. (e) Psychological stress This variable uses the categories 12 to 25 from the Kessler (K5) score that indicates whether a person has high or very high psychological distress. This measure is based on the psychological assessment questionnaire provided to the NATSISS surveyor to monitor nonspecific psychological stress of the surveyed person. In Australia, the measure has commonly been used in public mental health services and more recently in various health surveys (Sunderland et al 2011). (f) Positive social and emotional wellbeing This is one of only two continuous variables being estimated. This variable combines different indicators based on how the surveyed person felt in the last 4 weeks. This includes the feeling of calm and peaceful (SEWQ12), happy (SEWQ13), full of life (SEWQ14) and having a lot of energy (SEWQ15). (g) Household members ran out of money for basic living expenses This binary variable indicates whether household members ran out of money for basic living expenses. The question was asked twice in the questionnaire, first based on the condition in last 12 months( HHFSQ4) and second based on condition in last 2 weeks (NOMON2W). (h) Feelings of safety walking alone in the local area after dark Unlike the other variables, this variable is estimated using two steps. We first flag those who are walking alone after dark, and then look at who felt safe among those who were walking alone after dark. The original variable measuring feelings of safety actually contains five values: very safe, safe, neither safe nor unsafe, unsafe and very unsafe. The unsafe and very unsafe were combined to get those who were not feeling safe walking alone after dark. 12

13 (i) Whether victim of physical/threatened violence Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. The variable is a combination of two variables on the NATSISS with a value of one assigned to those who have either been a victim of physical violence (CJVQ1) or were threatened with physical violence in last 12 months. (j) Personally experienced stressor This variable identifies those who are experiencing different stressors in last 12 months (WSTROWN). There are around 24 types of stressor used for this variable including bad illness, bad accident, marriage, pregnancy, divorce or separation, death of family member or close friend, lost or change job, alcohol or drug related problems, abuse or violent crime and discrimination. (k) Identifies with clan, tribal or language group For this variable, the reference person in the family is asked whether he or she and the family identify themselves into a certain tribal group, language group, clan, mission or regional group Benchmarking variables between the Census and NATSISS The following are the variables that are not only available in both Census and NATSISS but also defined in a similar fashion so they can be benchmarked. (a) Age of person This variable in the Census indicates the age of a person based on the last birthday before the August Census collection date. The date of birth information provides the best estimate of age if it is available. Alternatively, stated age will be used only if date of birth is not provided. In the NATSISS we use AGEC. This variable is based on the age stated by the reference person. Similar to the age in Census, the age variable in NATSISS is mostly in single years. However, on the NATSISS the ABS confidentialises those who are older than 64 years old. We have therefore had to aggregate the age ranges to match between the two sources. (b) Sex of a person SEXP identifies the sex of a person. In both Census and NATSISS, there are only two options in this category, male or female, and if there is no answer on the NATSISS then the sex variable is imputed. (c) Relationship in household The relationship of a particular person within the household is another variable available on both the Census and the NATSISS. On the Census, the variable (RLHP) contains 29 different relationship types including family and non-family member with categories such as Husband, Wife in a registered marriage, Partner in a de facto marriage, Lone parent, Natural or adopted child, Step child, and so on. The variable (RLHHLC) on the NATSISS contains only eight groups husband, wife or partner; lone parent; child under 15; dependent student; non- 13

14 dependent child; other related individual; non-family member and lone person. We have therefore aggregated the Census classifications to match the survey classifications. (d) Level of highest non-school qualification The Non-School Qualification on the census (QALLP) describes the level of education of the highest completed qualification after secondary school. Although the definition of the matching variable in the NATSSIS is practically the same, the classification is slightly different and regrouping on both databases was needed to produce the same classification across these two databases. The not applicable category in both databases includes people with no qualification, people still studying for a first qualification and people aged under 15 years. (e) School status This variable indicates whether a person is attending school on a full time or part time basis. The variable on the Census (STUP) is not exactly the same as the matching variable on the NATSISS (FTPTSTDY) as the variable on the NATSISS is only applied to those aged 15 years and over while in the census, STUP is applied to everyone. Therefore, we have adjusted the Census variable to represent only those who are 15 years and above. (f) Labour force status Although the two datasets use a slightly different definition, the difference between the labour force status classification on the Census and the NATSISS is not great. On the Census, the variable LFSP indicates the person s labour force status in the week before the Census night, while on the NATSISS (EMPSTAC) the reference period was the entire week, Monday through Sunday, prior to the interview. (g) Occupation in main job In both the Census and the NATSISS, the occupation variable shows the occupation of any employed person classified using the new Australian New Zealand Standard Classification of Occupations (ANZSCO). The classification includes Managers, Professionals, Technicians and Trades Workers, Community and Personal Service Workers, Clerical and Administrative Workers, Sales Workers, Machinery Operators and Drivers and Labourers. (h) Equivalised household gross weekly income Income plays in important role in wellbeing. The variable Total Household Income (weekly) in the Census data (HIND) is calculated by summing the personal incomes reported by all household members aged 15 years and over. The Census collects personal income in ranges. Therefore, a specific dollar amount needs to be allocated to each person using median incomes for each range based on data from the Survey of Income and Housing. Household income is not calculated where a household member aged 15 years and over did not state their income, or was temporarily absent. These households are coded to the 'Partial income stated' category. The definition of the Household gross weekly income variable (WINCPH) in NATSISS is similar but it is presented as single dollar value rather than in income groups. This study further calculates equivalised household gross weekly income by adjusting the household income by the number of people in the household. This adjustment is necessary to compare the incomes of households with a different number of people in them. The 14

15 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. adjustment uses the 'modified OECD' equivalence scale, which is built up by allocating 1 point to the first adult, 0.5 points to each additional person who is 15 years and over, and 0.3 to each child under the age of 15 (Atkinson et al, 1995). (i) Family type This variable classifies families into different types of families. There are different categorisations of family type on the Census. The main category only contains four types of family: couple without children, couple with children, lone parent and other family while the most detailed breakdown has 17 categories. Family type on the NATSISS contains 13 categories that do not exactly match the categories in the Census. Therefore, the family type in the NATSISS was grouped to match the more general four categories of family type in the Census. (j) Tenure type and Landlord type The variable for tenure type is TEND in the Census and HSTENUC in the NATSISS. The variable describes whether a dwelling is owned, being purchased or rented. The landlord type variable (LLDD on the Census and TYPRENC on the NATSISS) records the landlord type of rented dwellings. Although not exactly the same, the categories from each data source are similar. For example, the dwelling categorised as being purchased on the Census is comparable with owner with a mortgage on the NATSSIS. This is also the case with landlord type. Moreover, the landlord type is aggregated further to be public (rented from State or Territory housing authority), private or other. (k) Monthly mortgage repayments On the 2011 Census, the value of mortgage repayments only is available (MRED), which matches the NATSISS definition. One difference is that the NATSISS variable is calculated on a weekly basis while the Census data is monthly. We have converted the weekly amount to monthly amounts. (l) Working motor vehicles owned by household members There is a slight difference in definitions in the motor vehicle benchmark. The variable on the Census (VEHD) refers to the number of registered motor vehicles owned or used by household members garaged or parked at or near a private dwelling on Census Night. On the NATSISS, the variable (NCARHHC) is defined as the number of working motor vehicles owned by household members. (m) Computers connected to the Internet in the household The variable on the NATSISS (HHITHQ6) asks whether any computers are connected to the Internet in the household, while the variable on the Census (NEDD) records the type of Internet connection most frequently used in addition to whether the dwelling has an Internet connection. For this model we have just used whether the household has or doesn t have an internet connection. 15

16 3 Building the Synthetic Unit Record data 3.1 Creation of the unit record data based on Census tables For spatial microsimulation, a survey dataset is required that provides the unit records. Traditionally, we would use ABS Confidentialised Unit Record Files for this, but there is not one available for the NATSISS, so we had to create one. The first step in the construction of the database was to create some empty observations. The empty observations contain only three variables: an observation number or person identifier, the geographic area from the Census and an initial weight. To get this weight, we need to know how many Indigenous people live in different areas of Australia. Therefore, the first thing we needed to do was decide which of the ASGC areas to use for this part of the analysis. This choice of area will affect the accuracy of our final reweighting, as knowing approximately where someone lives gives a starting point for the estimation process. The geographic area chosen for the first part of this estimation procedure was the Major Statistical Region from the ASGC structure. There are 14 regions in this classification: the Capital City and the balance of the State of the five larger States, namely, New South Wales, Victoria, Queensland, South Australia and Western Australia and one area each for Tasmania, Northern Territory, the Australian Capital Territory and Other Territories. In this project, we excluded the Other territory region which has a very small population. Based on the number of Indigenous people in the area from the Census, each region was populated with synthetic people. Note that this does not mean that the number of synthetic people in a region will match the actual number of people in that region as the number of observations in places like Sydney would be too large for the model to work efficiently. In creating the synthetic people, each synthetic person receives a weight to represent a number of people in the area. The initial weight is calculated as the actual number of people in the area from the Census divided by the number of synthetic people. After this synthetic database was created, six variables were imputed to each of the observations. These six variables were age, sex, labour force status, gross weekly household income, gross weekly household equivalised income, and family composition. For this project, age was reclassified into eight age groups: 0 to 14, 15 to 24, 25 to 34, 35 to 54, 55 to 64, 65 to 74, 75 to 84, and 85 years and over; and the labour force status has been simplified into employed, unemployed, not in the labour force, not applicable, and not stated. Not applicable applies to people who are not old enough to be employed (so aged under 15). For family composition, the classifications were couple family with no children, couple family with children, one parent family and other family. The not applicable classification was also used for those living in lone person households and group households. To allocate these six variables to the records in our synthetic database, we have used a probability table from a cross tabulation of these six variables from the Census data. Overall, there were 126 thousand categories allocated to our population. The use of only six variables for the first imputation was based on limitations of the Census Tablebuilder program. The next variable imputed was Individual gross weekly Income. This variable shows the income level of people aged 15 years and over in income ranges. There were 15 different 16

17 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. categories with 10 valid income brackets and the categories negative income, nil income, not stated, not applicable and overseas visitor. To impute the Individual gross weekly income bracket for each person we used the probability of the person being in a particular income group based on five other variables that have been imputed previously age, sex, labour force status, gross weekly household income and gross weekly household equivalised income. We then imputed two variables about the number of people in the family the number of dependent children in the family and the number of people in the family. The variable for the number of dependent children in the family gives the count of children under 15 years of age, or dependent students aged years, in the family. Although this can include up to three dependent children who were temporarily absent on Census night, the variable itself has a limitation. It provides an accurate count for up to five children and is then top coded, so the final category is six or more children. Unlike the number of dependent children in the family, the number of people in the family includes other related individuals who are not part of the nuclear family, such as in-laws, grandparents, uncles and so on. These two variables were imputed together because we could get more accurate estimates using a combination of equivalised income and non-equivalised income as the main variables in the probability table. Besides those two variables, we also used sex and family composition as the basis for the joint probability. After imputing these variables, we then imputed the tenure type. This variable indicates whether the dwelling the household lives in is a rental property or owned. This variable is only available for occupied private dwellings. The classification used for this variable is fully owned, being purchased, rented, other tenure, not stated, and not applicable. To impute this variable we used a probability table based on five variables age, sex, family composition, gross weekly household income and gross weekly household equivalised income. In estimating this variable, we started to find that there were some observations that did not fit into the characteristics shown in the probability table, so we decided to drop these observations from our synthetic dataset. The reason these observations may occur is that the imputation may assign a certain condition to an observation where, in the census, the condition does not exist in the area. For example, the modelling may impute a family with children with an income of $100 - $200 per week when no family with children earning this much exists in the area. We then estimated the number of people usually resident in the dwelling. Unfortunately, due to limitations in Table Builder, we had to drop the tenure variable and use the number of dependent children in the family, the number of people in the family, household equivalised income and household income in the probability table. All these variables combined should give a sense of how many people actually live in the household as the equalivisation of Household Income is based on the number of adults and children in the household. The next indicator imputed consists of two variables combined together the highest year of school completed, and the highest completed non-school qualification. These combine to form a level of education. The highest year of schooling completed shows the highest level of primary or secondary schooling completed and has categories of year 12 or equivalent, year 11 or equivalent, year 10 or equivalent, year 9 or equivalent, year 8 or below and did not go to school. People aged below 15 years are not applicable for this category. The highest completed non-school qualification includes postgraduate degree level, graduate 17

18 diploma and graduate certificate level, bachelor degree level, advanced diploma and diploma level as well as certificate level. These two variables are strongly correlated to each other since only a very few people have a Bachelor degree and above qualification without having finished year 12. Personal income, sex, age and labour force status are variables that are closely correlated with these education variables, and hence are used in the probability table. The next variable imputed was the relationship in household. This variable describes the relationship of each person in the household to the household reference person. This includes family and non-family member with categories such as husband, wife in a registered marriage, partner in a de facto marriage, lone parent, natural or adopted child, step child, foster child, grandchild, unrelated child, brother/sister, father/mother, cousin, uncle/aunt, nephew/niece and so on. The variables used to construct the probability table for this variable were age, sex, family composition, labour force status and personal income. After imputing the relationship variable, we imputed a combined variable which consisted of landlord type and weekly rent. These two variables give a specific attribute to those living in a rented dwelling. The landlord type variable shows the landlord type of the rented property categorised as private tenure, State or Territory housing authority (public) and other tenure. The rent is divided into 16 brackets from zero to $49 a week to 550 a week and over. The variables used to distribute the probability of renting are the number of people usually resident in the dwelling, the number of dependent children in the family, tenure type and gross weekly household income. The next variable imputed was the monthly housing loan repayment. This variable was only attributed to those households who were purchasing their dwelling. There are 18 payment brackets associated with this variable, and the probability has been estimated based on the distribution of the number of people usually resident in the dwelling, the number of dependent children in the family, age, tenure type and gross weekly household income. The next two variables imputed were dwelling related variables, being the type of internet connection and the number of motor vehicles. The type of internet connection variable shows the most frequently used type of Internet connection (separating broadband, dial-up and no connection) in a dwelling. The variables used for the probability table for the imputation of this variable were number of people usually resident in the dwelling, the count of dependent children in the family, age, tenure type and gross weekly household income. The number of motor vehicles variable shows the number of registered motor vehicles, including company owned vehicles, owned or used by household members. This variable is only applicable to occupied private dwellings. The variables used for the probability table to impute this variable were number of people usually resident in the dwelling, the number of dependent children in the family, age, tenure type and gross weekly household income. We then imputed two variables related to employment: occupation and industry of employment. The variables used to distribute the probability of each person s occupation were age, sex, qualifications, labour force status and individual gross weekly income. The industry of employment variable shows the Industry that a person is employed in based on the Australian and New Zealand Standard Industrial Classification (ANZSIC) 2006 classification. The ABS classifies a person s industry of employment based on the description of the business, and the main goods produced, or main services provided in the place the 18

19 Error! No text of specified style in document. Error! No text of specified style in document., Error! No text of specified style in document. person works. In addition, the name of the business, the employed person's occupation and main tasks and duties may also be used as additional information to determine the industry. As the occupation of the person determines the industry of employment, occupation is one of the variables used for the industry of employment probability matrix. Other variables used were age, sex, qualifications and individual gross weekly Income. In addition to the variables above, we also imputed the full time/part time student status and the usual address on census night indicator. This Full Time/Part Time student status variable is imputed because we wanted information on the school retention rate among Indigenous students. The variables used to construct the probability table for this variable were age, sex, family composition, labour force status and household equivalised income. The usual address on Census night indicator shows whether the person actually lives in the dwelling on a daily basis. This variable is needed because we have based our imputation on enumerated persons, so this variable identifies those living at their usual address. The imputation used age, sex, labour force status, personal income and the relationship in household in the probability tables. Table 1 summarises this imputation process, providing a summary of all the variables imputed and the variables used for the conditional probabilities. 19

20 Table 2 The sequences of imputation and the variables used for the conditional probability imputation step Variables Imputed Conditional Probability Basis 1 age, sex, labour force status, gross weekly household income, gross weekly household equivalised income, and family composition age, sex, labour force status, gross weekly household income, gross weekly household equivalised income, and family composition 2 individual gross weekly income age, sex, labour force status, gross weekly household income and gross weekly household equivalised income 3 the number of dependent children in the family and the number of people in the family equivalised income, non equivalised income, sex and family composition 4 tenure type age, sex, family composition, gross weekly household income and gross weekly household equivalised income 5 number of people usually resident in the dwelling 6 the highest year of school completed, and the non-school qualification: level of education Number of dependent children in the family, number of people in the family, household equivalised income and household income personal income, sex, age and labour force status 7 relationship in household age, sex, family composition, labour force status and personal income 8 landlord type and weekly rent number of people usually resident in the dwelling, number of dependent children in the family, tenure type and gross weekly household income 9 monthly housing loan repayment number of people usually resident in the dwelling, number of dependent children in the family, age, tenure type and gross weekly household income 10 type of internet connection number of people usually resident in the dwelling, number of dependent children in the family, age, tenure type and gross weekly household income 11 number of motor vehicles number of people usually resident in the dwelling, number of dependent children in the family, age, tenure type and gross weekly household income 12 occupation age, sex, qualifications, labour force status and individual gross weekly income 13 industry of employment occupation, age, sex, qualifications and individual gross weekly income 14 full-time/part-time student status age, sex, family composition, labour force status and household equivalised income The synthetic database that we have created using this process may not produce a set of observations that exactly match the Census cross-tabulations. The next step was to apply a reweighting method to make the current synthetic database much closer to the original Census tables. Before doing this, we decided to look at the differences between tables calculated from the synthetic dataset and census data for several of these variables before 20